Milling machine fault detection and identification based on a novel vitality index and temporal-residual network

This paper presents a novel technique for diagnosing faults in milling machines based on a new health indicator called the vitality index and a temporal-residual network (T-ResNet). Health monitoring of milling machines is crucial for ensuring production reliability and minimizing downtime, and acoustic emission (AE) signals offer a sensitive and effective means for early fault detection in this context. However, traditional features extracted from the acoustic emission signal, such as root mean square and kurtosis, are widely used for fault detection, but their sensitivity is compromised by noise. Furthermore, acoustic emission hit (AEH) features have received little attention in milling machine diagnosis. Therefore, in this study, AE signals are acquired from a milling machine for extraction of AEH features. The signals include both continuous and burst type AEHs. Burst-type hits are distinct due to their clear rise and decay, whereas continuous hits lack such differentiation from background noise. A new adaptive thresholding technique is used to effectively extract AEH features. The new technique adapts itself to the AE signal and takes into account both types of AEHs. The distribution of those features varies with the milling machine’s health status. To identify a distribution change, the Mann-Whitney test is applied to the features at multiple scales to obtain a new health indicator called the vitality index (VI). The vitality index changes significantly as the milling machine transitions from a normal operating state to a faulty condition. To identify the defect in the milling machine, the index is classified using a T-ResNet. The proposed method is validated using real-world industrial milling machine data and demonstrates fault detection superior to existing techniques.